Graph neural networks (GNNs) are crucial for graph node classification, yet existing interpretability methods focus mainly on instance-level or model-level analysis, neglecting the impact of nodes during training. This results in incomplete explanations, particularly regarding how other nodes influence a target node’s classification. To address this, we propose a novel interpretable method that optimizes graph classification tasks. Our approach provides insights into both the training and prediction phases by analyzing the cumulative impact of nodes on classification results. It uses counterfactual memory and attribution functions to measure the total influence of a node throughout both phases, offering a deeper understanding of node classification. Additionally, we examine how the long-tail distribution of node degrees affects GNN interpretability, showing that low-degree nodes are more sensitive to the influence of other nodes during training. This paper introduces new methods to explain node classification, incorporating both the training process and the long-tail distribution of node features.